Morphology and spectral sensitivity of long visual fibers and lamina monopolar cells in the butterfly Papilio xuthus

Extensive analysis of the flower‐visiting behavior of a butterfly, Papilio xuthus, has indicated complex interaction between chromatic, achromatic, and motion cues. Their eyes are spectrally rich with six classes of photoreceptors, respectively sensitive in the ultraviolet, violet, blue, green, red, and broad‐band wavelength regions. Here, we studied the anatomy and physiology of photoreceptors and second‐order neurons of P. xuthus, focusing on their spectral sensitivities and projection terminals to address where the early visual integration takes place. We thus found the ultraviolet, violet, and blue photoreceptors and all second‐order neurons terminate in the distal region of the second optic ganglion, the medulla. We identified five types of second‐order neurons based on the arborization in the first optic ganglion, the lamina, and the shape of the medulla terminals. Their spectral sensitivity is independent of the morphological types but reflects the combination of pre‐synaptic photoreceptors. The results indicate that the distal medulla is the most plausible region for early visual integration.

background is low, even if a sufficient degree of chromatic contrast exists (Koshitaka et al., 2011).Such an achromatic intensity contrast is generally crucial for detecting motion (Kaiser & Liske, 1974;Yamaguchi et al., 2008), but P. xuthus' motion vision is also sensitive to chromatic contrast (Stewart et al., 2015).
As a part of elucidating the mechanism underlying the synergy of achromatic and chromatic information, we have been studying the anatomy of how the spectral PRs and second-order visual interneurons may interact.In P. xuthus, the variety of PRs is housed in the array of ommatidia, each containing nine PRs (R1-R9; Figure 1b) in three fixed combinations, making their eyes a hexagonal patchwork of three spectrally distinct types of ommatidia.The combinations are UV/nB/dG/R (type I), V/sG/BB (type II), and wB/dG (type III).The shortwavelength-sensitive PRs (UV, V, nB, wB) are either R1 or R2, while the long-wavelength-sensitive PRs (dG, sG, R, BB) are assigned to R3-R8, and presumably to R9 (Figure 1c) (Arikawa, 2003;Arikawa & Uchiyama, 1996).The axons of all nine PRs are bundled at the ommatidium base and together enter the first optic ganglion, the lamina.In the lamina, they form a module called "cartridge," with five second-order neurons, the lamina monopolar cells (LMCs), tentatively termed L1-L5 (Figure 1b) (Matsushita et al., 2022).The axons of R1 and R2 leave the lamina with those of LMCs and run into the first optic chiasm.These PRs are called long visual fibers (LVFs) because their long axons reach the second optic ganglion, the medulla.R3-R9 PRs are short visual fibers (SVFs) terminating in the lamina.
The axons of LVFs and LMCs derived from a lamina cartridge form a "column" in the distal part of the medulla and terminate there.In the fruit fly, Drosophila melanogaster, the distal medulla is where the chromatic and (achromatic) motion information processing starts (Heath et al., 2020;Li et al., 2021;Pagni et al., 2021;Schnaitmann et al., 2018;Takemura et al., 2017).Assuming it is also the case in P. xuthus, detailed anatomy of LMCs and LVFs in the medulla is crucial to understand the mechanism underlying the complex visual signal processing.
Here, we studied the structure and function of LVFs and LMCs by combining single-cell electrophysiology, intracellular dye injection, and three-dimensional electron microscopy.We particularly focused on the correlation of their morphology in the lamina and medulla with their spectral sensitivities.

Animals
We used adults of the Japanese yellow swallowtail, P. xuthus, from a laboratory culture derived from females caught around the SOK-ENDAI campus in Kanagawa, Japan.The females laid eggs on fresh citrus leaves in the laboratory.We fed the hatched larvae with citrus leaves at 25 • C under the short-light (10L:14D) or long-light (14L:10D) cycle, respectively, resulting in diapausing or nondiapausing pupae.We kept newly emerged adults at 10 • C, feeding them with 20% sucrose solution once a week.We used adults aged 4-76 days after emergence for the experiments.

Electrophysiology
We recorded spectral responses of LVFs and LMCs by intracellular electrophysiology using glass micropipettes filled with Neurobiotin (Vector Laboratories)-containing KCl or K-acetate solution.
With its legs and wings removed, we mounted a butterfly on a stage and inserted a piece of chloridized silver wire into the base of an antenna as the reference electrode.First, we made a small cut in the first abdominal segment to suppress the pumping of the body fluid.
Next, we positioned the targeted eye at the center of the perimeter device with an optical fiber attached.The optical fiber's tip provided a point light source with a visual subtend angle of about 0.7 • , allowing single ommatidium stimulation (Takeuchi et al., 2006).When targeting LVFs, we inserted the electrode vertically into the retina through a hole on the dorsal cornea of the left or right eye (Figure 1d).For LMCs, we tilted the head around 35 • to the left to insert the microelectrode into the left lamina or the first optic chiasm (Figure 1b) through a hole opened beside the eye's dorsal edge (Figure 1e).The electrode resistance was 80-220 MΩ.
The sampling frequency was 4 samples/ms.For recording spectral responses, we used 21 narrowband interference filters and an electrical shutter to produce monochromatic light flashes of 30-ms duration at 1-s intervals.The other setup used for LMC recording consisted of an SEC-05X amplifier (Npi Electronic) and a Micro1401 A/D converter (CED) coupled with the WinWCP software (v4.0.5;J. Dempster, University of Strathclyde).The sampling frequency was set at 10 samples/ms.The stimulus was also a series of 21 monochromatic flashes, but using an LED array (365-691 nm [Belušič et al., 2016]) to allow averaging based on multiple quick scans of 50-ms flashes with 100ms intervals.In either case, we adjusted the lights' maximal intensity at 5 × 10 11 photon/cm 2 /s at the corneal surface (Chen et al., 2020).
The exit pupil of the stimulus light with a 1-mm diameter was placed 16 cm away from the eye.The subtended visual angle was thus about 0.7 • , smaller than P. xuthus' ommatidial acceptance angle, 1.9 • , and the interommatidial angle, 1.0 • (Takeuchi et al., 2006).

Light microscopy
We visualized the morphology of penetrated cells as follows.For PRs, we performed the following when they were LVFs (UV, V, B receptors).
Immediately after the electrophysiological measurements, we applied a positive current of 0.9-1.0nA for 3-5 min to inject Neurobiotin.The first DAB-stain protocol was applied to all LVFs and some LMCs.
First, the fixed brain sample was embedded in a gelatin/albumin mixture and further fixed overnight at 4 were incubated at 4 • C for 3 days in a secondary antibody solution: 1:300 Cy5-conjugated goat anti-mouse IgG (A10524; Life Technologies), 1:1000 Cy3-conjugated streptavidin (Jackson ImmunoResearch Laboratories), 1% NGS, and 0.5% Triton X-100 in 0.1 M PBS.After washing with 0.1 M PB (4 × 20 min), the slices set on gelatincoated slides were dehydrated, infiltrated, and mounted as described above.
For the third resin preparation, we washed the brains with 0.1 M PB (5 × 20 min) and incubated them at 4 • C for 5 days in 0.1 M PBS containing 1:1000 Cy3-conjugated streptavidin plus 0.5% Triton X-100.
We observed the DAB-stained samples using a BX60 light microscope (Olympus).We observed the fluorescent samples using a Nikon A1 laser confocal microscope (Nikon) with 560-and 638-nm lasers, the latter for the double staining, acquiring a stack of optical sections on the NIS-Elements software (Nikon).

Response analysis
We quantified the amplitude of each hyperpolarizing response to reveal the spectral properties of LMCs based on electrophysiology.
The LMCs start to respond at least about 5 ms after the stimulus onset (Chen et al., 2020).Therefore, we defined the response amplitude as the membrane potential difference between the 10-sample (1 ms) average immediately after the stimulus onset and the 10-sample average around the negative peak detected 10-60 ms after the stimulus onset; all measurements peaked within this time frame in sensitive wavelength ranges.We averaged amplitudes from five to eight trials and normalized the maximum to obtain relative spectral responses.Chen et al. (2020) reported spectral-opponent LMCs with depolarization in specific wavelength ranges, but we did not obtain single-LMC staining in these cases (Figure S1).
We first categorized LMC responses according to the ommatidia they potentially belong to.For this, we performed a hierarchical cluster analysis with Ward's method (Ward Jr., 1963), using response values in UV (365, 372, and 386 nm) and red (588, 600, 629, 651, 668, and 691 nm) wavelengths.
Next, we hypothesized that LMC spectral responses in the same ommatidia are related to LMC morphological types or other possible explanatory factors.We thus constructed linear regression models (Table 1) and compared their appropriateness with an information criterion.The relative response to the ith stimulus wavelength (R i , i ∈ {1, 2, . . .21}) is defined as where the function f n represents a normal distribution with two parameters, mean and standard deviation (SD).Mean's α i and β i and SD's σ i are random variables differing by stimulus wavelengths.The explanatory variable X in the mean represents one of the factors considered (Table 1): Model 1, LMC morphological types (0: L1, L2, or L5; 1: L3); Model 2, recording azimuth (degree from the eye's front); Model 3, recording elevation (degree from the eye's equator); Model 4, the maximum response amplitude (mV) among the trial-averaged spectral responses; Model 5, sex (0: male, 1: female); Model 6, seasonal form (0: spring, 1: summer); Model 7, postemergence age (day).Dummy variables, 0 or 1, expressed qualitative characters in Models 1, 5, and 6, so the parameter β i represents the difference between the two categories.In Model 1, we grouped the LMCs into two, L1/2/5 versus L3, because L1, L2, and L5 receive inputs mainly from G-sensitive R3 and R4 receptors, while the latter receives inputs from R1 and R2 short-wavelength receptors (Matsushita et al., 2022).We did not consider L4 because we did not encounter L4s.As for a case without any explanatory variables, we excluded the member β i X in Model 0.
We estimated parameters in a Bayesian framework via the Markov chain Monte Carlo simulation using CmdStan (Stan Development Team, 2022;v2.29.2).Five hundred samples were obtained from four chains; in each, every fourth generation was sampled in 2000 iterations after a warm-up of 500, with the target acceptance range of 0.8.We checked the parameters' convergence by trace plots and the potential scale reduction factor (Rhat) less than 1.1.Although the mean value ideally ranges from 0 through 1 for a normalized response, we did not assume logistic regression or truncated distribution models because of the parameters' poor convergence.Instead, we compared the appropriateness among the linear regression models using the widely applicable information criterion (WAIC) (Watanabe, 2009).a For AIC scaling (Gelman et al., 2013), the WAIC difference to the best model was multiplied by twice the number of total measured responses.b Dummy variable, 0 or 1, was applied.

Electron microscopy
We identified the detailed morphology of LVFs and LMCs using serial block-face scanning electron microscopy (SBF-SEM) followed by manual segmentation.
We dissected the compound eyes with lamina and medulla attached in the prefixative containing 2.5% glutaraldehyde and 2% paraformaldehyde dissolved in 0.15 M sodium cacodylate (pH 7.3, CB) with 2 mM CaCl 2 (Ca-CB).After fixing the eyes for about 30 min on ice, we removed the cornea and processed the tissues overnight at 4 • C in the same fixative.We then washed the tissues in Ca-CB for 5 h, embedded the samples in gelatin, and sectioned them at 100-μm thickness using a vibration microtome.We sequentially processed the sections as follows, interleaved with washes (5 × 3 min) with Milli-Q water (QW): 2% osmium tetroxide (OsO 4 ) plus 1.5% K-ferrocyanide in Ca-CB for 1 h on ice, 1% thiocarbohydrazide in QW for 20 min at RT, 2% OsO 4 in QW for 30 min at RT, 1% uranyl acetate in QW overnight at 4 • C, and Walton's lead aspartate solution for 30 min at 60 • C. We then dehydrated the sections with an ethanol series, infiltrated them with acetone, and embedded them in hard Spurr's resin (Polysciences, Inc.).
We prepared another sample (sample 2), where we followed the LVFs and LMCs from the lamina to the medulla through the first optic chiasm.This was possible only at the eye's center, where the LVFs and LMCs axons run straight down to the medulla.We collected 12,050 images at 50-nm intervals (602.5 μm).Each image covered 80 × 80 μm 2 (10,000 × 10,000 pixels, pixel size = 8.00 nm) from the basal retina to the distal 100 μm of the first optic chiasm or 80 × 120 μm 2 (10,000 × 15,000 pixels, pixel size = 8.00 nm) more proximally for tracing LVFs and LMCs in 12 lamina cartridges/medulla columns.
For three-dimensional reconstruction of LVFs and LMCs, we selected 1906 images at 100-nm intervals to cover the entire lamina (∼90 μm) and the distal 100 μm of the first optic chiasm, 664 images at 200-nm intervals to trace the axons in the chiasm's middle part, 492 images at 100-nm intervals to continuously trace the axons to the medulla's distal edge, and 1200 images at 100 nm intervals to cover the medulla layers 1-5a (see Figure S1).We separately aligned the four image stacks using the StackReg plugin for the ImageJ/FIJI software.We used the TrakEM2 plugin to perform manual segmentation.

Morphometrics
We focused on the morphology of spectrally different LVFs to correlate the medulla columns with the lamina cartridges and eventually with the ommatidial types.We quantified the arborization of LVF medulla terminals to evaluate if the morphology is diagnostic for the spectral classes.Using the BX60 light microscope equipped with camera lucida (Olympus), we traced the DAB-stained LVFs to compare the number of processes among different spectral classes (Welch's t-test).
We extended the diagnosis three dimensionally using SBF-SEM images.We developed an automatic protocol to count the endpoints' number of the traced LVF medulla terminals by modifying York's Matlab script (York et al., 2018) to skeletonize our high-resolution binary image stacks.In advance, we extracted 600 continuous images (42 μm) from the most proximal part of each LVF and filled enclosed spaces and invaginations in the volume to prevent abnormal skeletonization.
Because the algorithm also detected many extra branches in thick parts, we added an automatic step to remove endpoints, defined by volume erosion (expanding 0-valued pixels by 6-pixel radius) followed by dilation (expanding 1-valued pixels by 13-pixel radius) of the eroded volume.In addition, we altered a skeletonization threshold from 100 to 7 so as not to miss the distal ends of thin processes.Finally, we compared the number of endpoints among different LVF medulla terminals to see if this count effectively estimates the spectral classes.

Morphology of LVFs
We stained nine narrow-blue (nB)-, nine UV-, eight violet (V)-, and six wide-blue (wB)-sensitive LVFs using Neurobiotin injection: the nB and UV cells are from type I ommatidia, V are from type II, and wB are from type III (Figure 1c). Figure 2 shows representative spectral responses of these LVFs recorded in the retina (Figure 2a) with their traced images in the lamina and medulla (Figure 2b).We traced the LVFs in horizontal sections to follow the axons running in the first optic chiasm more easily, but this orientation obscured the lateral processes that extended vertically.The V receptors have processes most extensively in the lamina, which matches our previous observations (Takemura et al., 2005).At the medulla's distal surface, the LVF axons bend, followed by a smooth region, and have a brush-like terminal in layer 4b (Figure 2b) (Hamanaka et al., 2013).We manually counted the lateral processes in the traced images and found that the UV and V receptors have more processes than the nB and wB receptors (p < .01 by Welch's t-test), with an overlap at around 17 (Figure 2c).The spectral responses and traced images of cells in Figure 2c are shown in Figure S2.
Second, we observed LVFs' medulla terminals in the SBF-SEM sample 1, attempting to identify ommatidial types without having the morphology in the lamina and their spectral sensitivities.We counted the endpoints of 20 LVFs paired in 10 medulla columns while rotating the traced images (Figure 3). Figure 3a shows a terminal (R1 of column 1, 1-R1) viewed in three directions; black dots indicate the automatically detected endpoints.We confirmed that the automatic endpoint detection, ranging from 30 to 93 in the count, agreed with the manual detection.Figure 3b shows the endpoint numbers of five LVF pairs with the accommodating ommatidial types identifiable (Matsushita et al., 2022;Takemura & Arikawa, 2006).Supposedly, the large difference in column 1's pair (1-R1 vs. 1-R2) corresponds to terminals belonging to a type I ommatidium with UV/nB receptors (cf.Figures 1c and 2c).
Figure 3c shows the other five LVF pairs with the ommatidial types ambiguous.We also quantified the variance of endpoint positions in x-, y-, and z-axes and found no obvious sign of columnar heterogeneity (Figure S3).As found in the light-microscopic count, these pairs probably come around the overlapping range between the UV/V and nB/wB classes (hatched range in Figure 2c).We also note that the LVF axons are thin in layers 1-3 but are slightly thicker and rougher in layers 4a and 4b.

Morphology of LMCs
We could follow 12 modules (lamina cartridges/medulla columns) through the first optic chiasm in the SBF-SEM sample 2, where we found two LVFs and five LMCs in each module.We manually segmented all LVFs and LMCs in three cartridges/columns a (Figure 4a), b, and c (Figure S4).The transverse section of the lamina cartridge is rhomboidal (Figure 4b).In comparison, the medulla column appears roundish with the axons of LVFs and LMCs closely packed (Figure 4c, see also the rotated images of cell bundle in Figure 4a).In the lamina, the LMCs have extensive lateral processes as already described (Matsushita et al., 2022).The LMCs have bushy terminals whose shape and depth are characteristic of the LMC type.In the previous study (Matsushita et al., 2022), we tentatively termed the LMCs as L1-L5 based on the morphology in the lamina and the preliminary tracing result in the medulla.Here, we completed the tracing in the medulla, confirming the terminal structures in three columns and dye injection coupled with electrophysiology (see next section).We thus conclusively identify L1-L5 based on the terminal morphology.
The L1s form a round bush fitting in the medulla layer 1, occupying the most distal region of the column.Some processes reach the middle of layer 2 in columns a and c.
The L2s have a roundish but flatter bush terminating in the distal layer 2, immediately proximal to the L1's terminal.The bush further extends some processes that end in layer 3, while one reaches layer 4b.The long process has knobby swellings at layer 4a's distal region.In addition, the L2 axon may have a few short processes at the medulla's distal edge.
The L3s have a club-shaped bush terminating in layer 3. The axon bears several processes around the boundary of layers 1 and 2, and a few short processes in the proximal layer 2. We found no processes going deeper.
The L4s' bush stretches in the entire layer 3, with some intrusions in the distal layer 4, and a thin vertical process reaches layer 4b or 5.The bush is composed of two clusters surrounding the distal and proximal parts of the L3's club-shaped terminal.Each cluster is round as in L1 and L2, but may be somewhat disrupted (see column b).The central axon is smooth in layers 1 and 2.
L5s are peculiar, branching into some vertical processes at the medulla's distal border.One of the branches penetrates an adjacent column, while others stay in the home column.The branches continue down at least to layer 5.Each branch has sparse bushes in layer 2, a thickened region around the border between layers 3 and 4, and some knobs in layer 5.In columns a and c, we found further bifurcation in layer 5 and a short branch into another neighboring column from layer 1 to layer 3 in column a and down to layer 4b in column c.
Overall, we unambiguously identified five LMCs, L1-L5.We have detected no sign of morphological specificity of LMCs related to the spectral types of ommatidia.melanogaster's visual system (Takemura et al., 2008).In the medulla, we detected inter-LVF (Figure 4d at layer 4b) and inter-LMC (Figures 4e   and 4f at layers 1 and 3, respectively) connections but could not find evidence that LVFs connect to LMCs.

Spectral responses of LMCs
We recorded the spectral responses of LMCs, stained them in 164 individuals (Figure 1e), and selected 46 LMCs for further analysis (Figure S5).The rejected cases include those with multiple LMCs stained equally dense (Figure S1a) and two "spectral-opponent" LMCs (Chen et al., 2020) where PRs are co-stained (Figure S1b).Forty-six accepted cells are five L1s, five L2s, 34 L3s, and two L5s: we failed to identify L4s electrophysiologically.
Figure 5 shows representative dye-filled L1, L2, L3, and L5 samples.As described above, the bush positions in the medulla indicate the LMC types.We identified the medulla layers based on the antisynapsin staining or the tissue's background fluorescence (Hamanaka et al., 2013).Axon bending indicates the position of the medulla's distal surface (arrowheads in Figure 5a-d).Figure 5a shows an L1, terminating in the medulla layer 1.The round bush in Figure 5b   (588-691 nm) ranges.The analysis resolved three clusters, classes I, II, and III, at 2.6 of the "height" scale of the dendrogram (dotted line i in Figure 6a).These classes appear to be reasonable for the ommatidial types, represented by the composite of each type's PR spectral sensitivities (Chen et al., 2020).The largest cluster, class I, comprises 32 samples with responses of more than .5 at both 372 and 600 nm.
Class I is further divided into two subclasses at the height of 2.0 (dotted line ii in Figure 6a): 18 samples with higher responses at R wavelengths (Figure 6b) and 14 with lower responses (Figure 6c).Class II consists of four samples with a sensitivity less than .5 at 372 nm (Figure 6d).
Class III consists of 10 samples with a sensitivity less than .5 at 600 nm (Figure 6e).None showed responses smaller than .5 at both 372 and 600 nm.
We also performed the same analysis with all 21 stimulus wavelengths, including the middle wavelength range.The analysis reveals that the class III counterpart contains two samples with large responses in red wavelengths (Figure S6, ID 13 and 44), which disagrees with the feature of the ommatidial type III, which lacks R receptors' inputs (Chen et al., 2020).We thus follow the UV-and R-based clustering shown in Figure 6a for subsequent statistics.
We executed linear regression model analyses to elucidate the factor(s) contributing to dividing class I LMCs into two subclasses (Figure 6a-c).For 32 LMCs of class I, we analyzed their responses at 629, 651, and 668 nm (i = 18, 19, and 20 in Equation 1; 96 responses in total) using eight models with different explanatory variables (Table 1).
The WAIC evaluation showed that sex (Model 5) best explained the intra-type variation (Figure S7), while no explanatory variable (Model 0) behaved similarly.Other variables, including the morphological type of LMCs (L1/2/5 or L3), were ranked lower.The rank appears to match that 12 among 14 low red subclass LMCs were from females (Figure 6a).

Morphological identification of medulla column types
The eye of P. xuthus is a collection of three spectral types of ommatidia.
They are randomly arranged in the hexagonal array, which is retained in the array of the lamina cartridges and of the medulla columns.
The ommatidial spectral heterogeneity is identifiable among lamina cartridges based on the PRs' arborization, which is specific to their spectral sensitivities.For example, the UV-and V-sensitive LVFs have long lateral processes penetrating the neighboring cartridges, while no processes of B-sensitive LVFs go out from the home cartridges (Takemura et al., 2005).
Here, we studied whether or not the medulla columns were also identifiable by the LVFs' terminal structures.Dye injection following electrophysiological identification of spectral sensitivity suggested that the medulla terminals of the UV and V receptors appeared bushier than those of B receptors (Figures 2 and S2).However, more precise anatomy at the electron microscopic level revealed that the number of branches is not necessarily an explicit feature of the LVFs' spectral identity (Figure 3).bushier than wild-type individuals.In contrast, the class III and IV da neurons develop simpler dendritic arbors (Kim et al., 2006).In the pupal eye disc of D. melanogaster, Ss is expressed in a subset of dR7 (Drosophila-R7 photoreceptor) precursors, which eventually express UV-sensitive rhodopsin, Rh4.On the other hand, the Ss-negative dR7s express another UV-sensitive rhodopsin, Rh3 (Wernet et al., 2006).In P. xuthus, Ss is expressed selectively in B receptor precursors, eventually expressing B-sensitive rhodopsin, but not in those of the UV and V receptors, both expressing UV-sensitive rhodopsin in the adult (Kitamoto et al., 2000).The loss-of-function ss mutation made all ommatidia into type II where both LVFs (R1 and R2) were V receptors expressing UV-rhodopsin (Perry et al., 2016).Presumably, the Ss promotes B-sensitive rhodopsin expression and develops the medulla terminal with fewer arbors in P. xuthus.
In addition to LVFs, we traced the LMCs to see whether their branching patterns correspond to the ommatidial type.However, we have found such features neither in the lamina nor the medulla (Figures 4   and S4).Therefore, we are still unable to identify the ommatidial types corresponding to any given medulla columns solely based on the morphology of the LVF and LMC terminals.

Species comparison of the medulla circuit
A single ommatidium of butterflies and flies bears nine (R1-R9) and eight (dR1-dR8) PRs, respectively.Despite the difference in PR numbers, both have two LVFs in each ommatidium (R1 and R2 in P. xuthus; dR7 and dR8 in D. melanogaster).However, the LVF pairs of P. xuthus and D. melanogaster have different morphological characteristics and developmental origins.Morphologically, the fly dR7 and dR8, respectively, occupy the distal and proximal regions of the central rhabdomere in the retina (Hardie, 1985).On the other hand, the butterfly R1 and R2 both contribute to the rhabdom in its distal tier (Arikawa, 2003).The terminals in the medulla appear to retain the organization: in flies, dR7s terminate more distally than dR8s (Fischbach & Dittrich, 1989), while the butterfly R1 and R2 terminate at the same depth (Figure 4).At the molecular level, the butterfly R1 and R2 are both equivalent to the fly dR7, expressing Prospero (Pros) (Cook et al., 2003;Perry et al., 2016).Interestingly, the duplicated dR7-like butterfly PR is recruited in the same position where a "mystery cell" has been described in D.
melanogaster (Tomlinson et al., 1987).The mystery cell, a PR precursor, disappears during development; therefore, flies may have secondarily lost another dR7.The P. xuthus' PR cells equivalent to the fly dR8 are R9s; both express Senseless (Sens) (Frankfort et al., 2001;Perry et al., 2016): R9s are basal PRs contributing to the rhabdom at its most proximal region near the basement membrane (Figure 1b,c).However, the butterfly R9s are SVFs (Matsushita et al., 2022), and it remains unknown whether the butterfly R9s secondarily lost the long axon or the fly dR8 secondarily acquired the long axon.
The medulla circuit involved in wavelength information processing has been extensively studied in D. melanogaster.The dR7 and dR8 mutually inhibit directly and also via a distal medulla amacrine neuron 9 (Dm9), then feed the Dm8 neuron in the same column to construct spectral opponency (Heath et al., 2020;Kind et al., 2021;Li et al., 2021;Schnaitmann et al., 2020;Takemura et al., 2015).The mutual inhibition of R1 and R2 via histamine-gated chloride channel also happens in P. xuthus, where the medulla circuit is still unknown.In the medulla, we detected inter-LVF (Figure 4d at layer 4b) and inter-LMC (Figures 4e and 4f at layers 1 and 3, respectively) connections but could not find evidence that LVFs connect to LMCs.The LVF-LMC connections may exist outside the medulla as in D. melanogaster (Kind et al., 2021), but our present analysis is restricted to the distal part of the medulla neuropil.As in flies, the medulla terminals of

Variability of LMCs and their spectral sensitivities
We unambiguously identified five LMCs per cartridge/column of P.
xuthus in two complete series of electron microscopy image stacks: only careful segmentation of all LMCs allows such a convincing statement.Another example analyzed at this level is D. melanogaster, where a single lamina cartridge also contains five anatomically distinct LMCs (Fischbach & Dittrich, 1989) with two possible additional subtypes (Rivera-Alba et al., 2011).However, whether the L1-L5 of P. xuthus and D. melanogaster are homologous or not remains unknown.
We also presume that the LMCs were evolutionarily homologous among species, although an extensive molecular analysis is required to convincingly demonstrate their "deep" homology (Ozel et al., 2021).
Nevertheless, we tentatively termed the P. xuthus LMCs partially based on the nomenclature of four LMCs of Papilio aegeus, PaL1-PaL4 (Ribi, 1987).The P. aegeus' L1 (PaL1) densely bears short processes within the home cartridge in the lamina and terminates with a round bush at the most distal layer 1 in the medulla.These features are common to those of P. xuthus' L1 (Figures 4a and S4).The PaL2 is similar to P. xuthus' L2, with long processes penetrating the adjacent cartridges and the medulla terminal at layer 2. The medulla terminals of PaL3 and P. xuthus' L3 share the club-like bushes at layer 3.However, PaL3 bears only a few processes within the home cartridge in the lamina, unlike P. xuthus' L3.No LMC type in our study resembled PaL4, characterized by processes only in the proximal layer of the lamina and a tapering terminal down to the medulla layer 3. Of L4 and L5 in P. xuthus, the latter may be homologous to PaL4, given the extension of lamina processes across several cartridges and a sparse bush in the distal medulla with a proximal elongation.The response characteristics of L4s remain unknown because we did not encounter L4s in the present electrophysiology.A possible reason is that the main shaft of L4s is narrow and escaped from the electrode tip.Or, L4s may depolarize, which could be a basic nature of the neuron receiving hyperpolarizing inputs from the presynaptic L3s (Matsushita et al., 2022).We thus injected Neurobiotin when we encountered LMC-like quick (Chen et al., 2020) but depolarizing responses.However, most of such trials co-labeled SVFs and LMCs.
According to the previous connectome analysis in the lamina (Matsushita et al., 2022), L1, L2, and L3 receive major inputs from all PRs in the home cartridge with some variations; L1 and L2 receive more inputs from R3 and R4 G receptors, while L3 receives inputs mainly from R1 and R2 LVFs.In contrast, L5s receive a few inputs from the home cartridge's R3 and R4 G receptors but more inputs from cells in adjacent cartridges.The simulation of LMCs' spectral sensitivities based on the synapse counts has revealed that L1, L2, and L3 in a given ommatidial type exhibit similar sensitivity profiles, indicating LMC's morphological types may have little effect on spectral sensitivities.
Pronounced differences become evident when we compare one cell type, for example, L1, among three ommatidial types.The difference is attributable to the ommatidial spectral heterogeneity (Figure 1c), which served as the basis of the present cluster analysis (Figure 6).
The distinction between the two subclasses of class I cells (Figure 6b,c) is independent of the LMCs' morphological types.
Although slight, the sample's sex provided the best in the WAIC evaluation (Table 1); class I cells with low red sensitivity are biased toward females, attributable to any possible sexual dimorphism in the visual system.At the PR level, sexual dimorphism of spectral sensitivity has been first identified in pierid butterflies.In the small white Pieris rapae, the spectral sensitivity of V receptors is narrower in males because of the male-specific filter pigment (Arikawa et al., 2005).The R receptor of the eastern pale clouded yellow Colias erate peaks at 640 nm, but the females in addition have 620-nm-and 660-nm-peaking R receptors: the variability is also due to the sexually dimorphic reddish filter pigments (Ogawa et al., 2013).The sexual dimorphism of the eye's spectral organization is now evident in lycaenids (Sison-Mangus et al., 2006) and nymphalids (Ilić et al., 2022;McCulloch et al., 2016), indicating that such a phenomenon is more common among butterflies than previously thought.Although nothing is known yet in papilionids, there could still be unknown dimorphism in P. xuthus' visual system, even at the level of synaptic interaction of PRs and LMCs, that could back up the present results.

Perspectives
Extensive analysis of the visual behavior of foraging P. xuthus has revealed complex interaction between the chromatic, achromatic, and motion cues (Kinoshita & Arikawa, 2014;Koshitaka et al., 2011;Stewart et al., 2015).The present functional anatomy of LVFs and LMCs provides a clue toward understanding the mechanism underlying the parallel processing of various visual signals.
The UV-B-G-R tetrachromacy indicates that both LVFs (UV and B channels) and SVFs (G and R channels) contribute to seeing colors.
Assuming that the chromatic information processing starts in the distal medulla, the SVFs' signal should, most likely, be relayed to the medulla via LMCs.The spectral sensitivities of L1-L3 are similar in a module corresponding to an ommatidium but differ among modules connected to different ommatidial types (Figure 6) (Matsushita et al., 2022).It is also possible that the negative components of the spectral opponent LVFs feed long-wavelength information to the medulla circuit (Chen et al., 2020).
The LMCs' spectral variability among ommatidial types could be the source of chromatic contrast contributing to motion vision.For motion vision, achromatic input plays a major role in flies and bees whose SVFs all exhibit identical spectral sensitivity (Hardie, 1985;Wakakuwa et al., 2005).The fly's LMCs almost exclusively receive inputs from the spectrally identical SVFs in flies (Rivera-Alba et al., 2011) and feed the achromatic motion vision circuit (Takemura et al., 2013;Yamaguchi et al., 2008).Assuming that the LMCs' primary func-tion is to control motion vision also in P. xuthus, the spectrally variable SVFs (Figure 1) add the chromatic property to their motion vision system (Stewart et al., 2015), which presumably has increased the fitness of the color-dependent behavior in butterflies (van der Kooi et al., 2021).In addition, a search of chromatic contrast-dependent motionsensitive neurons in P. xuthus has identified large neurons (Céchetto et al., 2022) having the processes in the medulla layers 1-3 (Céchetto et al., in preparation), where L1-L4 have their terminals (Figure 4).
The present anatomy of PRs and LMCs in the medulla makes a step-by-step analysis of their connections toward higher order neurons feasible.For example, the amacrine-like Dm8 and Dm9 neurons are core members of the chromatic processing circuit in the distal medulla in D. melanogaster (Heath et al., 2020;Pagni et al., 2021).These neurons cover multiple columns (Takemura et al., 2015), so the analysis of P. xuthus' counterparts will provide insights into how inter-ommatidial interaction plays a role.Such information would eventually contribute to understand the algorithm creating a series of color-coding neurons in the mushroom body (Kinoshita & Stewart, 2022), a higher brain region responsible for learning and memory that serve the basis of P. xuthus' sophisticated color vision properties (Kinoshita & Arikawa, 2023).
Following a diffusion time of 20-60 min at room temperature (RT), we isolated the head, removed the posterior cuticle facing to the thorax, and immersed the sample in a fixative containing 4% paraformaldehyde, 0.25% glutaraldehyde, and 0.2% picric acid in 0.1 M phosphate buffer (PB, pH 7.4) overnight at 4 • C.After removing the head capsule and cornea, we processed the sample according to one of the following three protocols: (1) 3,3′-diaminobenzidine (DAB) staining to visualize neuronal arborizations with a high signal-noise ratio, (2) double fluorescent staining with anti-synapsin to superimpose the neuron's image in the medulla layers, and (3) fluorescent staining with resin embedding to keep the original morphology as much as possible.

Figures 4
Figures 4 and S4 show the morphology of LVFs (R1/2) and LMCs (L1-L5) of three neighboring lamina cartridges and their corresponding medulla columns obtained from the SBF-SEM sample 2. We hereafter refer to the bundle shown in Figure 4 as cartridge/column a and those in Figure S4 as cartridges/columns b and c.Based on the previous anatomy of LVFs in the lamina(Matsushita et al., 2022;Takemura & Arikawa, 2006), we unambiguously identified cartridge a's R1 and b's R2 as UV receptors, which extended longish and fewer processes in the Figure 4d-f shows examples of synapselike structures in the medulla judged from the distribution of synaptic vesicles; as in the lamina (Matsushita et al., 2022), we could not encounter any "T-bar"-like structures associated with synapses in D.
positions in layer 2, indicating the cell is an L2.Due to their thin lamina axons, L1 and L2 samples often have co-stained cells; Figure 5a,b contains faintly co-stained L3 terminals reaching layer 3. Figure 5c shows an L3, characterized by the club-shaped terminal in layer 3 and the thick (4-5 μm in diameter) axon at the medulla's distal edge.Figure 5d is an L5, showing a thin shaft, long lateral processes in the lamina, and a bifurcated axon in the medulla terminating in layer 5.The recording traces shown below (Figure 5e-h) are spectral responses from the LMCs in Figure 5a-d, respectively.Figure 6 shows the normalized spectral response functions of all 46 LMCs analyzed with the result of statistical clustering.Given that LMC sensitivities to UV and R wavelengths are diagnostic for three spectral classes reflecting the type of the home ommatidium (Chen et al., 2020), we performed a cluster analysis of the responses in the UV (365-386 nm) and R (b) The number of endpoints on long visual fibers' (LVFs) medulla terminals automatically counted on three-dimensional volumes, based on serial block-face scanning electron microscopy (SBF-SEM) sample 1.(a) The terminal of R1 of column 1 (1-R1) viewed from three directions, that is, viewed horizontally (0 • ), vertically (90 • ), and from the bottom.Black dots indicate endpoints.(b) Samples with accommodating ommatidial types identifiable.Numbers indicate the column ID.(c) Samples with the accommodating ommatidial types are ambiguous, showing that the endpoint count is not always sufficient to identify the LVF's spectral class.
Morphology of long visual fibers (LVFs) (R1 and R2) and lamina monopolar cells (LMCs) (L1-L5) in the lamina (La) and the medulla (Me), determined based on serial block-face scanning electron microscopy (SBF-SEM) sample 2. (a) Three-dimensional reconstruction of a lamina cartridge and the corresponding medulla column, viewed horizontally (0 • ) and vertically (90 • ).Individual cells are taken from the 90 • view.Three short double lines in the L5 image denote processes' further elongation.(b-f) Electron microscopy images taken at the levels indicated by arrowheads with b-f in panel (a).R1, R2, and L1-L5 are colored.(b) Lamina cartridge (dashed curve).(c) Medulla column (dashed curve) at the distal border of the medulla.(d-f) Synapse-like structures (white arrows) from R1 to R2 in layer 4b (d), L1 to L3 in layer 1 (e), and L3 to L4 in layer 3 (f).
Morphology and spectral responses of four lamina monopolar cells (LMCs).(a-d) Fluorescent staining of four LMCs in the lamina (La) and the medulla (Me), viewed in horizontal (0 • ) and vertical (90 • ) directions.Axon bending points (arrowheads) indicate the distal surface of the medulla.LMC morphological types were identified based on Figures 4a and S4.(a) L1, characterized by the round bush terminal in the medulla layer 1.The weakly co-stained is the L3 in the same module.(b) L2, characterized by the long lateral processes in the lamina (arrows) and the round bush terminal in the medulla layer 2. L3 was slightly co-stained.(c) L3.(d) L5. (e-h) Raw responses to 21 monochromatic flashes from 365 to 691 nm of the samples shown in panels (a-d).
Evaluation of linear regression models for "class I" lamina monopolar cell (LMC) spectral responses at 629, 651, and 668 nm in Papilio xuthus.
TA B L E 1 P. xuthusLVFs have numerous outputs whose targets are unidentified.
cal extension of the present study is to elucidate whether the P. xuthus medulla column contains two Dm8 neurons corresponding to two dR7like LVFs, R1 and R2.In addition, the morphology of still unidentified neurons may allow us to distinguish spectral types among medulla columns.